Domain Transfer Research Articles

Digital twin-assisted dual transfer: A novel information-model adaptation method for rolling bearing fault diagnosis

Rolling bearing fault diagnosis is of great importance to the safety management of mechanical equipment. The scarcity of labelled fault data makes it difficult to adequately perform the training process of intelligent diagnosis models, and this will result in these intelligent models not being effectively and widely used in practice. Although some recent studies have verified that the addition of dynamic model response to the training process will greatly improve the ability of the model with low cost and high efficiency, it is still stuck in poor effect caused by large information distribution difference between dynamic model response and real measured data. Focusing on this issue, a digital twin-assisted dual transfer (DTa-DT) method with information and model adaptation was proposed for rolling bearing fault diagnosis. Different from the traditional digital-analogue driven transfer methods, the proposed DTa-DT aims to simultaneously synthesize data information transfer and feature model transfer together with domain transfer error minimization. In particular, it should be noted that the DTa-DT architecture consists of a dual transfer learning process, including digital twin-driven information transfer (DTd-IT) and digital-analogue-driven model transfer (DAd-MT), where the information is collaborated with the model to improve the integrated transfer diagnosis effect under sampling. On one aspect, with the employment of bearing dynamic model responses, DTd-IT is innovatively designed to establish the transfer of dynamic information and measured information. The information distribution difference between these twin data and real measured data is effetely adjusted with the introduced actual inference components, where the twin data with low information distribution difference can be well fusion generated by the information transfer digital twin (ITDT) model. On the other aspect, considering the truth that there are still small sample cases of real measured data and information distribution differences will affect the quality of the twin data, a digital-analogue driven model transfer (DAd-MT) method is further proposed, where the deep branch transfer network (DBTN) model with improved convolutional neural network (CNN) is used to achieve an accurate fault diagnosis effect with the help of digital twin data. Experiments and wear analysis verified that the proposed DTa-DT can significantly reduce the distribution difference between the dynamic model response and the real measured data, thus achieving low-cost and efficient rolling bearing transfer diagnosis compared to other ten state-of-the-art deep learning models. It can be predicted that the proposed dual transfer architecture provides more opportunities for the practical application of intelligent fault diagnosis under small sample sizes.

A Diffusion Model Translator for Efficient Image-to-Image Translation.

Applying diffusion models to image-to-image translation (I2I) has recently received increasing attention due to its practical applications. Previous attempts inject information from the source image into each denoising step for an iterative refinement, thus resulting in a time-consuming implementation. We propose an efficient method that equips a diffusion model with a lightweight translator, dubbed a Diffusion Model Translator (DMT), to accomplish I2I. Specifically, we first offer theoretical justification that in employing the pioneering DDPM work for the I2I task, it is both feasible and sufficient to transfer the distribution from one domain to another only at some intermediate step. We further observe that the translation performance highly depends on the chosen timestep for domain transfer, and therefore propose a practical strategy to automatically select an appropriate timestep for a given task. We evaluate our approach on a range of I2I applications, including image stylization, image colorization, segmentation to image, and sketch to image, to validate its efficacy and general utility. The comparisons show that our DMT surpasses existing methods in both quality and efficiency. Code will be made publicly available.

A Survey of Label-Efficient Deep Learning for 3D Point Clouds.

In the past decade, deep neural networks have achieved significant progress in point cloud learning. However, collecting large-scale precisely-annotated point clouds is extremely laborious and expensive, which hinders the scalability of existing point cloud datasets and poses a bottleneck for efficient exploration of point cloud data in various tasks and applications. Label-efficient learning offers a promising solution by enabling effective deep network training with much-reduced annotation efforts. This paper presents the first comprehensive survey of label-efficient learning of point clouds. We address three critical questions in this emerging research field: i) the importance and urgency of label-efficient learning in point cloud processing, ii) the subfields it encompasses, and iii) the progress achieved in this area. To this end, we propose a taxonomy that organizes label-efficient learning methods based on the data prerequisites provided by different types of labels. We categorize four typical label-efficient learning approaches that significantly reduce point cloud annotation efforts: data augmentation, domain transfer learning, weakly-supervised learning, and pretrained foundation models. For each approach, we outline the problem setup and provide an extensive literature review that showcases relevant progress and challenges. Finally, we share our views on the current research challenges and potential future directions. A project associated with this survey has been built at https://github.com/xiaoaoran/3D label efficient learning.

Sim2Real: Generative AI to Enhance Photorealism through Domain Transfer with GAN and Seven-Chanel-360°-Paired-Images Dataset.

This work aims at providing a solution to data scarcity by allowing end users to generate new images while carefully controlling building shapes and environments. While Generative Adversarial Networks (GANs) are the most common network type for image generation tasks, recent studies have only focused on RGB-to-RGB domain transfer tasks. This study utilises a state-of-the-art GAN network for domain transfer that effectively transforms a multi-channel image from a 3D scene into a photorealistic image. It relies on a custom dataset that pairs 360° images from a simulated domain with corresponding 360° street views. The simulated domain includes depth, segmentation map, and surface normal (stored in seven-channel images), while the target domain is composed of photos from Paris. Samples come in pairs thanks to careful virtual camera positioning. To enhance the simulated images into photorealistic views, the generator is designed to preserve semantic information throughout the layers. The study concludes with photorealistic-generated samples from the city of Paris, along with strategies to further refine model performance. The output samples are realistic enough to be used to train and improve future AI models.

Multi-Log Grasping Using Reinforcement Learning and Virtual Visual Servoing

We explore multi-log grasping using reinforcement learning and virtual visual servoing for automated forwarding in a simulated environment. Automation of forest processes is a major challenge, and many techniques regarding robot control pose different challenges due to the unstructured and harsh outdoor environment. Grasping multiple logs involves various problems of dynamics and path planning, where understanding the interaction between the grapple, logs, terrain, and obstacles requires visual information. To address these challenges, we separate image segmentation from crane control and utilise a virtual camera to provide an image stream from reconstructed 3D data. We use Cartesian control to simplify domain transfer to real-world applications. Because log piles are static, visual servoing using a 3D reconstruction of the pile and its surroundings is equivalent to using real camera data until the point of grasping. This relaxes the limits on computational resources and time for the challenge of image segmentation, and allows for data collection in situations where the log piles are not occluded. The disadvantage is the lack of information during grasping. We demonstrate that this problem is manageable and present an agent that is 95% successful in picking one or several logs from challenging piles of 2–5 logs.

Unsupervised multilingual machine translation with pretrained cross-lingual encoders

Multilingual Neural Machine Translation (MNMT) has recently made great progress in training models that can translate between multiple languages. However, MNMT faces a significant challenge due to the lack of sufficient parallel corpora for all language pairs. Unsupervised machine translation methods, which utilize monolingual data, have emerged as a solution to this challenge. In this paper, we propose a method that leverages cross-lingual encoders, such as XLM-R, in an unsupervised manner (i.e., using monolingual data and bilingual dictionaries) to train a MNMT model. Our method initializes the MNMT model with a pre-trained cross-lingual encoder and employs two levels of alignment to further align the representation space in MNMT model. Experimental results demonstrate that our method outperforms strong baseline systems and exhibits robust domain and language transfer capabilities while preserving the performance of the original pre-trained encoder on other downstream tasks.

Fabric Defect Detection Method Using SA-Pix2pix Network and Transfer Learning

This paper proposes a fabric defect detection algorithm based on the SA-Pix2pix network and transfer learning to address the issue of insufficient accuracy in detecting complex pattern fabric defects in scenarios with limited sample data. Its primary contribution lies in treating defects as disruptions to the fabric’s texture. It leverages a generative adversarial network to reconstruct defective images, restoring them to images of normal fabric texture. Subsequently, the reconstituted images are subjected to dissimilarity calculations against defective images, leading to image segmentation for the purpose of defect detection. This approach addresses the issues of poor defect image reconstruction accuracy due to the limited ability of remote dependency modeling within the generator’s convolutional neural network. It also tackles deficiencies in the generative adversarial network’s loss function in handling image details. To enhance the structure and loss function of the generative adversarial network, it introduces self-attention mechanisms, L1 loss, and an improved structural loss, thus mitigating the problems of low defect image reconstruction accuracy and insufficient image detail handling by the network. To counteract the issue of declining model training accuracy in the face of sparse complex fabric defect samples, a channel-wise domain transfer learning approach is introduced. This approach constrains the training of the target network through feature distribution, thereby overcoming the problem of target network overfitting caused by limited sample data. The study employs three methods to experimentally compare and investigate five distinct complex pattern fabric defects. The results demonstrate that, when compared to two other defect detection methods, the approach advocated in this paper exhibits superior detection accuracy in scenarios with limited sample data.

A Structure-Aware Framework of Unsupervised Cross-Modality Domain Adaptation via Frequency and Spatial Knowledge Distillation.

Unsupervised domain adaptation (UDA) aims to train a model on a labeled source domain and adapt it to an unlabeled target domain. In medical image segmentation field, most existing UDA methods rely on adversarial learning to address the domain gap between different image modalities. However, this process is complicated and inefficient. In this paper, we propose a simple yet effective UDA method based on both frequency and spatial domain transfer under a multi-teacher distillation framework. In the frequency domain, we introduce non-subsampled contourlet transform for identifying domain-invariant and domain-variant frequency components (DIFs and DVFs) and replace the DVFs of the source domain images with those of the target domain images while keeping the DIFs unchanged to narrow the domain gap. In the spatial domain, we propose a batch momentum update-based histogram matching strategy to minimize the domain-variant image style bias. Additionally, we further propose a dual contrastive learning module at both image and pixel levels to learn structure-related information. Our proposed method outperforms state-of-the-art methods on two cross-modality medical image segmentation datasets (cardiac and abdominal). Codes are avaliable at https://github.com/slliuEric/FSUDA.

Omni-Training: Bridging Pre-Training and Meta-Training for Few-Shot Learning.

Few-shot learning aims to fast adapt a deep model from a few examples. While pre-training and meta-training can create deep models powerful for few-shot generalization, we find that pre-training and meta-training focus respectively on cross-domain transferability and cross-task transferability, which restricts their data efficiency in the entangled settings of domain shift and task shift. We thus propose the Omni-Training framework to seamlessly bridge pre-training and meta-training for data-efficient few-shot learning. Our first contribution is a tri-flow Omni-Net architecture. Besides the joint representation flow, Omni-Net introduces two parallel flows for pre-training and meta-training, responsible for improving domain transferability and task transferability respectively. Omni-Net further coordinates the parallel flows by routing their representations via the joint-flow, enabling knowledge transfer across flows. Our second contribution is the Omni-Loss, which introduces a self-distillation strategy separately on the pre-training and meta-training objectives for boosting knowledge transfer throughout different training stages. Omni-Training is a general framework to accommodate many existing algorithms. Evaluations justify that our single framework consistently and clearly outperforms the individual state-of-the-art methods on both cross-task and cross-domain settings in a variety of classification, regression and reinforcement learning problems.

PatchAugNet: Patch feature augmentation-based heterogeneous point cloud place recognition in large-scale street scenes

Point Cloud Place Recognition (PCPR) in street scenes is an essential task in the fields of autonomous driving, robot navigation, and urban map updating. However, the domain gap between heterogeneous point clouds and the difficulty of feature characterization in large-scale complex street scenes pose significant challenges for existing PCPR methods. Most PCPR methods only take into account point clouds collected by the same platforms and sensors, thus they are with poor domain transferability. In this paper, we propose PatchAugNet, which utilizes patch feature augmentation and adaptive pyramid feature aggregation to achieve better performance and generalizability for Heterogeneous Point Cloud-based Place Recognition (HPCPR) tasks. Firstly, multi-scale local features are extracted by the pyramid feature extraction module. Secondly, local features are enhanced by the patch feature augmentation module to overcome the domain gap problem and achieve better feature representation as well as network generalizability. Finally, a global feature is generated using an adaptive pyramid feature aggregation module, which automatically adjusts and balances the proportion of intra-scale and inter-scale features according to the scene content. To evaluate the performance of PatchAugNet, a large-scale heterogeneous point cloud dataset consisting of high-precision Mobile Laser Scanning (MLS) point clouds and helmet-mounted Portable Laser Scanning (PLS) point clouds is collected. The dataset covers various street scenes with a length of over 20 km. The comprehensive experimental results indicate that PatchAugNet achieves State-Of-The-Art (SOTA) performance with 83.43 % recall@top1% and 60.34 % recall@top1 on unseen large-scale street scenes, outperforming existing SOTA PCPR methods by + 9.57 recall@top1% and + 15.50 recall@top1, while exhibiting better generalizability. For source code and detailed experimental results, please refer to: https://github.com/WHU-USI3DV/PatchAugNet.

Analysis on the Effect of Phase Noise on the Performance of Satellite Communication and Measurement System

An oscillator is a key component of a satellite communication measurement and control system, performing symmetry precisely as a time frequency reference. At the same time, the phase noise index has a close coupling relationship with the overall performance of the entire system, while persistently breaking the symmetry property of the oscillator during work. It is very important to study and reasonably allocate the phase noise index. According to the theoretical formula of phase noise, this paper analyzes the power law spectral model in the frequency domain and the noise jitter characteristics in the time domain. Using the carrier tracking loop in the measurement system, the frequency domain transfer model of phase noise is established, and typical analysis results are given. A discrete fractional integration algorithm is proposed, which can generate the phase noise time domain sequence under the given power law spectral model coefficients. The proposed algorithm is more realistic compared with the previous numerical calculation method, and has sufficient accuracy compared with the results of the instrument. After frequency domain conversion, the RMS deviation between the simulated noise sequence in the frequency domain and the measured single sideband power spectral density is less than 2.5 dB, indicating that the phase noise sequence can reflect the frequency domain characteristics more completely. A communication measurement simulation system is built, and a discrete sequence simulation analysis method combining frequency domain and time domain is provided, and the coupling relationship of key indicators such as phase noise, thermal noise, communication data rate, modulation method and bit error rate is synthesized. The results show that the BER of the QPSK/BPSK communication system will not be significantly reduced if the phase jitter RMS caused by the phase noise is less than 5 degrees, so 5 degrees can be used as a reference for the decomposition of the carrier SSB phase noise index. The simulation results have been successfully applied to a satellite inter-satellite link system, which has universal practical application value.

Lattice Boltzmann simulations on transient radiative transfer problems in irregular geometries with constant or graded refractive index

This work is concerned with the transient radiative transport in inhomogeneous two- or three-dimensional semitransparent graded-index media with irregular geometry. The lattice Boltzmann method is applied to the graded-index media mapped with non-orthogonal grids to solve the radiative transfer while the finite volume method expresses the angular processing. The solution of the transient radiation from the methodology developed is firstly validated with good accuracy by considering benchmark solutions available in the literature for comparisons for various forms of geometry including a cube, elliptic cylinder as well as irregular shapes. Afterwards, the solver is used to analyze the radiative transfer in graded-index media including regular and irregular geometries, and various governing parameters such as laser and optical properties, scattering laws, incident angle of the incoming beam, refractive index, and temperature distributions. Suitable calculations of the transient radiometric heat flux and incident energy help to evaluate the time-resolved reflectance and transmittance. The results from this study indicate that the present methodology and boundary treatments are simple and suitable to deal with multidimensional radiative transfer in irregular domains.

A Review on Cross Weather Traffic Scene Understanding Using Transfer Learning for Intelligent Transport System

Intelligent transport systems (ITS) have revolutionized the transportation industry by integrating cutting-edge technologies to enhance road safety, reduce traffic congestion and optimize the transportation network. Scene understanding is a critical component of ITS that enables real-time decision-making by interpreting the environment's contextual information. However, achieving accurate scene understanding requires vast amounts of labeled data, which can be costly and time-consuming. It is quite challenging to Understand traffic scene captured from vehicle mounted cameras. In recent times, the combination of road scene-graph representations and graph learning techniques has demonstrated superior performance compared to cutting-edge deep learning methods across various tasks such as action classification, risk assessment, and collision prediction. It's a grueling problem due to large variations under different illumination conditions. Transfer learning is a promising approach to address this challenge. Transfer learning involves leveraging pre-trained deep learning models on large-scale datasets to develop efficient models for new tasks with limited data. In the context of ITS, transfer learning can enable accurate scene understanding with less data by reusing learned features from other domains.
 This paper presents a comprehensive overview of the application of transfer learning for scene understanding in cross domain. It highlights the benefits of transfer learning for ITS and presents various transfer learning techniques used for scene understanding. This survey paper provides systematic review on cross domain outdoor scene understanding and transfer learning approaches from different perspective, presents information on current state of art and significant methods in choosing the right transfer learning model for specific scene understanding applications.

3D-2D Distance Maps Conversion Enhances Classification of Craniosynostosis.

Diagnosis of craniosynostosis using photogrammetric 3D surface scans is a promising radiation-free alternative to traditional computed tomography. We propose a 3D surface scan to 2D distance map conversion enabling the usage of the first convolutional neural networks (CNNs)-based classification of craniosynostosis. Benefits of using 2D images include preserving patient anonymity, enabling data augmentation during training, and a strong under-sampling of the 3D surface with good classification performance. The proposed distance maps sample 2D images from 3D surface scans using a coordinate transformation, ray casting, and distance extraction. We introduce a CNNbased classification pipeline and compare our classifier to alternative approaches on a dataset of 496 patients. We investigate into low-resolution sampling, data augmentation, and attribution mapping. Resnet18 outperformed alternative classifiers on our dataset with an F1-score of 0.964 and an accuracy of 98.4 %. Data augmentation on 2D distance maps increased performance for all classifiers. Under-sampling allowed 256-fold computation reduction during ray casting while retaining an F1-score of 0.92. Attribution maps showed high amplitudes on the frontal head. We demonstrated a versatile mapping approach to extract a 2D distance map from the 3D head geometry increasing classification performance, enabling data augmentation during training on 2D distance maps, and the usage of CNNs. We found that low-resolution images were sufficient for a good classification performance. Photogrammetric surface scans are a suitable craniosynostosis diagnosis tool for clinical practice. Domain transfer to computed tomography seems likely and can further contribute to reducing ionizing radiation exposure for infants.

Cross domain fault diagnosis method based on MLP-mixer network

The quality of rolling bearings determines the safety of mechanical equipment operation, and bearings with more precise structures are prone to damage due to excessive operation. Therefore, cross domain fault diagnosis of bearings has become a research hotspot. To better improve the accuracy of bearing cross domain fault diagnosis, this study proposes two models. One is a cross domain feature extraction model constructed using a mixed attention mechanism, which recognizes and extracts high-level features of bearing faults through channel attention and spatial attention mechanisms. The other is a bearing cross domain fault diagnosis model based on multi-layer perception mechanism. This model takes the feature signals collected by the attention mechanism model as input to identify and align the differences between the source and target domain features, facilitating cross domain transfer of features. The experimental results show that the mixed attention mechanism model has a maximum accuracy of 97.3 % for feature recognition of different faults, and can successfully recognize corresponding signal values. The multi-layer perception model can achieve the highest recognition accuracy of 99.5 % in bearing fault diagnosis, and it can reach a stable state when it iterates to 26, and the final stable loss value is 0.28. Therefore, the two models proposed in this study have good application value.

A verilog design for cascading a large-capacity asynchronous FIFO with a synchronous FIFO

In modern large scale ASIC designs, multiple clock systems are often involved, which can create problems with data transfer in different clock domains. A practical solution to this problem is the use of asynchronous FIFOs (First In First Out) for buffering the transfer of data from different clock domains. A high-capacity asynchronous FIFO cascaded with a synchronous FIFO is designed based on a conventional asynchronous FIFO using the Verilog language. The input data is processed across the clock domain by the asynchronous FIFO, and the data output from the asynchronous FIFO is cached and output again by the synchronous FIFO. The module increases the FIFO depth while enabling data to be transferred across the clock domain. The simulation is completed within the Modelsim software which accordingly verifies the two main roles of the FIFO in processing data, namely the cross-clock domain and the data caching function. Simulation results show that the asynchronous FIFO data is written and read correctly and that the empty/full flag signal is correct.

Progressive dual-domain-transfer cycleGAN for unsupervised MRI reconstruction

Supervised MRI reconstruction methods perform well when provided with matched undersampled and fully sampled data pairs. However, acquiring paired data can be expensive and impractical in several clinical settings, such as cine MRI. As a result, recent studies have focused on reducing the reliance on paired data. Yet, most unsupervised approaches have attempted to directly convert the undersampled data domain into the fully sampled data domain, potentially leading to subpar reconstruction performance due to significant domain discrepancies. In this study, we propose a progressive dual-domain-transfer cycleGAN (PDD-GAN) to effectively address this issue. Our proposed method develops a dual-domain framework in an unsupervised manner, enabling the learning of representations from both image and frequency domains. Simultaneously, we break down the direct domain transfer problem into a multi-stage issue and solve it progressively, correcting reconstruction errors and preserving anatomical information at each transfer step. We conduct comprehensive experiments demonstrating that our approach outperforms several state-of-the-art supervised and unsupervised models on two public datasets. The code is publicly available at https://github.com/Coolwen1997/PDD-GAN.

Activity classification using unsupervised domain transfer from body worn sensors

Activity classification has become a vital feature of wearable health tracking devices. As innovation in this field grows, wearable devices worn on different parts of the body are emerging. To perform activity classification on a new body location, labeled data corresponding to the new locations are generally required, but this is expensive to acquire. In this work, we present an innovative method to leverage an existing activity classifier, trained on Inertial Measurement Unit (IMU) data from a reference body location (the source domain), in order to perform activity classification on a new body location (the target domain) in an unsupervised way, i.e. without the need for classification labels at the new location. Specifically, given an IMU embedding model trained to perform activity classification at the source domain, we train an embedding model to perform activity classification at the target domain by replicating the embeddings at the source domain. This is achieved using simultaneous IMU measurements at the source and target domains. The replicated embeddings at the target domain are used by a classification model that has previously been trained on the source domain to perform activity classification at the target domain. We have evaluated the proposed methods on three activity classification datasets PAMAP2, MHealth, and Opportunity, yielding high F1 scores of 67.19%, 70.40% and 68.34%, respectively when the source domain is the wrist and the target domain is the torso.

Wildfire Detection Using Convolutional Neural Networks and PRISMA Hyperspectral Imagery: A Spatial-Spectral Analysis

The exacerbation of wildfires, attributed to the effects of climate change, presents substantial risks to ecological systems, infrastructure, and human well-being. In the context of the Sustainable Development Goals (SDGs), particularly those related to climate action, prioritizing the assessment and management of the occurrence and intensity of extensive wildfires is of utmost importance. In recent times, there has been a significant increase in the frequency and severity of widespread wildfires worldwide, affecting several locations, including Australia, Italy, and the United States of America. The presence of complex phenomena marked by limited predictability leads to significant negative impacts on biodiversity and human lives. The utilization of satellite-derived data with neural networks, such as convolutional neural networks (CNNs), is a potentially advantageous approach for augmenting the monitoring capabilities of wildfires. This research examines the generalization capability of four neural network models, namely the fully connected (FC), one-dimensional (1D) CNN, two-dimensional (2D) CNN, and three-dimensional (3D) CNN model. Each model’s performance, as measured by accuracy, recall, and F1 scores, is assessed through K-fold cross-validation. Subsequently, T-statistics and p-values are computed based on these metrics to conduct a statistical comparison among the different models, allowing us to quantify the degree of similarity or dissimilarity between them. By using training data from Australia and Sicily, the performances of the trained model are evaluated on the test dataset from Oregon. The results are promising, with cross-validation on the training dataset producing mean precision, recall, and F1 scores ranging between approximately 0.97 and 0.98. Especially, the fully connected model has superior generalization capabilities, whilst the 3D CNN offers more refined and less distorted classifications. However, certain issues, such as false fire detection and confusion between smoke and shadows, persist. The aforementioned methodologies offer significant perspectives on the capabilities of neural network technologies in supporting the detection and management of wildfires. These approaches address the crucial matter of domain transferability and the associated dependability of predictions in new regions. This study makes a valuable contribution to the ongoing efforts in climate change by assisting in monitoring and managing wildfires.

Source-Free Domain Adaptation via Target Prediction Distribution Searching

Existing Source-Free Domain Adaptation (SFDA) methods typically adopt the feature distribution alignment paradigm via mining auxiliary information (eg., pseudo-labelling, source domain data generation). However, they are largely limited due to that the auxiliary information is usually error-prone whilst lacking effective error-mitigation mechanisms. To overcome this fundamental limitation, in this paper we propose a novel Target Prediction Distribution Searching (TPDS) paradigm. Theoretically, we prove that in case of sufficient small distribution shift, the domain transfer error could be well bounded. To satisfy this condition, we introduce a flow of proxy distributions that facilitates the bridging of typically large distribution shift from the source domain to the target domain. This results in a progressive searching on the geodesic path where adjacent proxy distributions are regularized to have small shift so that the overall errors can be minimized. To account for the sequential correlation between proxy distributions, we develop a new pairwise alignment with category consistency algorithm for minimizing the adaptation errors. Specifically, a manifold geometry guided cross-distribution neighbour search is designed to detect the data pairs supporting the Wasserstein distance based shift measurement. Mutual information maximization is then adopted over these pairs for shift regularization. Extensive experiments on five challenging SFDA benchmarks show that our TPDS achieves new state-of-the-art performance. The code and datasets are available at https://github.com/tntek/TPDS.